Recording Employees on Video for AI Datasets: Notice, Consent and Audio Rules

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why workplace video needs more than the employer's permission

An employer can own the camera and the file, but it cannot waive the separate rights of the people in the frame. A clip of a technician rebuilding a pump carries at least four layers: the footage copyright, each worker's privacy and publicity interests, any recorded conversation, and biometric data if faces or voices can be measured. The layers are mapped in our guide to rights layers in a video clip. This page covers the people layer for physical work at operating businesses.

Computer-use capture (screen, keystrokes, browser telemetry) has its own monitoring problems, covered in collecting computer-use demonstrations at work. The decision on faces, consent versus anonymization, is in licensing video that contains faces. Here the focus is notice, releases, audio and the people around the worker.

Which states require notice before employers monitor workers

As of October 2026, a small number of US states require written notice before an employer electronically monitors employees, and the statutes differ in what counts as monitoring. Check the site's state before capture starts, not after footage arrives.

• Connecticut (Conn. Gen. Stat. 31-48d): employers must give prior written notice of the types of electronic monitoring and post it conspicuously. The state's definition reaches camera and photo-optical systems, so a capture rig on a production line is in scope [2].
• New York (Civil Rights Law 52-c): employers that monitor phone, email or internet access must give written notice at hiring, get a written or electronic acknowledgment, and post the notice [1]. Its text is aimed at communications monitoring, so whether a camera rig is covered is a question for counsel. Many employers give the notice anyway.
• Delaware (19 Del. C. 705): notice is required before monitoring phone, email or internet use, either as a one-time acknowledged notice or a daily electronic notice [3].

For a dataset buyer, a statutory monitoring notice is the floor and is not consent for AI training. A notice that says "cameras may be used for safety and quality" describes a monitoring purpose. It does not tell a worker that the footage will be licensed to a third party to train models. Ask for both documents: the monitoring notice and a separate, project-specific release. Existing safety-camera archives are a separate case, covered in workplace safety and near-miss video.

What a worker release for AI training should say

A usable release names AI and machine-learning training explicitly, names the recipient class, and states what happens if the worker withdraws. Generic photo and video releases signed for marketing or training films often predate AI use. Commercial stock libraries now add AI and ML training language to their model releases for that reason [9].

Check these clauses when you review a supplier's release form:

• Purpose: "training, evaluating and improving machine-learning and AI models," not only "internal training" (which a worker will read as employee training).
• Recipients: licensing to third parties, including model developers outside the employer.
• Media scope: video, audio, still frames, derived pose or skeleton data, transcripts, and any face or voice measurements.
• Voluntariness: a plain statement that declining or withdrawing has no effect on pay, scheduling, evaluation or employment.
• Withdrawal mechanics: how a worker withdraws, and what withdrawal means for clips already delivered (exclusion from future deliveries, removal from unshipped batches, and how licensees are notified).
• Identity handling: whether faces, name badges, tattoos and voices will be blurred, removed or kept.
• Compensation: if workers are paid to participate, the amount and that it is not conditioned on outcomes.

Voluntariness is the hard part in an employment relationship. EU data protection authorities have long held that employee consent is rarely freely given because of the imbalance of power [10]. US law has no exact equivalent, but the same fact pattern weakens any release. If a supervisor recruits the crew, if non-participants are moved to worse shifts, or if the only people filmed are those who could not refuse, treat the releases as weak. Ask how recruitment was run and how many workers declined.

When the audio track changes the legal analysis

Audio is often the riskiest part of a workplace clip, because recording a conversation can trigger wiretap and eavesdropping statutes that silent video does not. Federal law and most states allow recording with one party's consent. A minority of states, California among them, require the consent of every party to a confidential communication [4].

California Penal Code 632 makes it an offense to record a confidential communication with a recording device without the consent of all parties. A communication is confidential when circumstances reasonably indicate that a party wants it confined to the participants [4]. A head-mounted camera on a field technician can capture a customer explaining a billing dispute, or two coworkers talking about a medical appointment. The worker's release does not supply the customer's or the coworker's consent.

Practical options, in order of how much they reduce risk:

1. Strip audio at capture when the model only needs visual action (assembly, picking, tool handling). Configure the device with the microphone off, not muted in post.
2. Keep only machine and task audio by recording in areas with no expected conversation, and remove speech segments with voice activity detection before delivery. Keep the removal log.
3. Keep speech with all-party consent, signage, and a verbal notice at the start of customer interactions, as call centers do. The checks mirror call-recording consent for AI training.

If you need speech, ask whether a transcript alone would do, and whether speaker voices must be retained at all. Retaining raw voice also raises the voiceprint question below.

Biometric exposure when faces or voices can be measured

Ordinary video is not automatically biometric data, but face geometry or voiceprints extracted from it are. Illinois BIPA covers scans of face geometry and voiceprints, excludes photographs, and requires a written release and a published retention and destruction schedule [5]. Texas requires notice and consent before a biometric identifier is captured for a commercial purpose [6]. Washington regulates enrolling biometric identifiers for a commercial purpose [7].

The risk often arrives downstream. A supplier may deliver raw clips, but a curation pipeline that runs face detection, re-identification or speaker diarization may create face-geometry or voiceprint data on your side. The litigation over IBM's Diversity in Faces dataset shows how a research face dataset can draw BIPA claims years after collection [8]. For the state-by-state detail, see biometric data in AI training: BIPA, CUBI and Washington.

Two buyer controls help. Require the supplier to state whether any biometric template was computed before delivery. Then set an internal rule that face embeddings and speaker embeddings are not computed on the licensed set unless the releases cover it.

Customers, visitors and other bystanders in frame

Bystanders signed nothing, so the dataset plan has to keep them out of the frame or make them unidentifiable. In a retail backroom, a warehouse dock or a home-service visit, the camera will catch customers, delivery drivers, contractors and other employees who did not opt in.

• Frame control: egocentric (head- or chest-mounted) cameras point at hands and tools and catch fewer faces than fixed wide-angle cameras. That is one reason egocentric video of skilled manual work suits manual-task datasets.
• Exclusion zones: no capture in restrooms, locker rooms, break areas or anywhere workers change. Some state laws ban surveillance in such areas outright.
• Redaction: blur faces and readable screens, badges, license plates and paperwork of non-consenting people. Record the method, tool version and QA sample rate.
• Signage: posted notice at entrances where customers or visitors may be filmed, plus a stop procedure if someone objects.
• Minors and health settings: stop capture or route to specialist review. Clinical footage follows HIPAA de-identification (see surgical video datasets).

Unionized sites and collective bargaining

At a unionized site, introducing new cameras or a new use of existing footage may be a subject the employer has to bargain over, and the collective bargaining agreement may already restrict surveillance. Ask the supplier whether the site is covered by a CBA, whether it has a technology or monitoring clause, and whether the union was notified of the capture project. A written union acknowledgment, or counsel's memo that bargaining was not required, belongs in the diligence file. A filming project that starts a grievance can halt capture mid-collection and cast doubt on footage already shot.

Pre-capture consent checklist for buyers

Use this checklist before capture starts. For licensed archives, use it as a document request. Store the answers alongside the delivery manifest, and see consent and notice records for AI training data for how to audit them.

Illustrative example: invented to show structure; it does not describe an available dataset.

An illustrative per-clip consent record, as it might appear in a JSONL manifest:

Illustrative example: invented to show structure; it does not describe an available dataset.

{"clip_id": "site07_cam2_0418_0932", "worker_pseudonym": "W-0113", "release_version": "AI-REL-v3", "release_signed": "2026-03-02", "monitoring_notice_ack": "2026-02-27", "state": "CT", "audio": "removed_at_capture", "bystanders_redacted": true, "redaction_method": "face+badge blur, 10% manual QA", "biometric_templates_computed": false, "withdrawal_status": "active", "cba_site": false}

Pseudonymous worker IDs let you remove a withdrawn worker's clips without holding names. For delivery formats and manifests, see dataset delivery formats and schemas.

How SourceX handles workplace video requests

SourceX sources operational datasets from US companies on request, including new recordings of hands-on work, and manages licensing and ongoing purchases. It does not hold footage in stock, and a request does not guarantee a match. It does not source generic CCTV. Buyers describe the data they need, every release is approved by the supplying company, and each dataset is rights-reviewed for ownership and consents. Personal details are removed or replaced before delivery, with the method recorded and a sample checked; no method is perfect.

The process runs Find, Assess, Agree, Transact and Manage, and nothing is contracted until a supplier agrees. Delivery happens through private, access-controlled workflows only after an executed agreement. For the broader picture of who buys this footage, read do AI labs buy video of people working, the video data hub, and the consent management glossary entry. You can describe your workplace video requirements to SourceX.

Request consented workplace video from operating businesses

If you need video of real physical work with notice, releases and audio handling documented per dataset, describe the tasks, settings and capture constraints. SourceX looks for US businesses that hold or can record that data and prepares diligence materials on source, rights, preparation and allowed use for each dataset. Start a buyer request.